Patent Application
20140334260
The present disclosure relates to seismic data processing. More specifically, the present disclosure relates to neural network based mapping of extensions of hydraulic fracturing events during fluid injection and well production. Seismic data processing has long been associated with the exploration and development of subterranean resources such as hydrocarbon reservoirs.
Hydraulic fracturing can be used to increase conductivity of a subterranean formation for recovery or production of hydrocarbons and to permit injection of fluids into subterranean formation or into injection wells. In a typical hydraulic fracturing operation, a fracturing fluid is injected under pressure into the formation through a wellbore. Particulate material known as proppant may be added to the fracturing fluid and deposited in the fracture as the fracture is formed to hold open the fracture after hydraulic fracturing pressure is relaxed.
Microseismic waves are generated at the tip of propagating hydraulic fractures that can, if monitored, provide information about a front of the progressing fractures while injecting fluid into the reservoir to aid in avoiding environmental and production problems. Monitoring microseismic waves generated by propagating hydraulic fractures may present a challenge since the signal-to-noise ratio between microseismic events and background noise can be small, and acquisition systems used for such monitoring may have to record a huge amount of data.
Some hydraulic fracturing monitoring techniques are described in: R. D. Barree, “Application of pre-frac injection/falloff tests in fissured reservoirs field examples,” SPE paper 39932, presented at the 1998 SPE Rocky Mountain Regional Conference, Denver, Apr. 5-8, 1998; C. L. Cipolla and C. A. Wright, “State-of-the-art in hydraulic fracture diagnostics,” SPE paper 64434, presented at the SPE Asia Pacific Oil and Gas Conference and Exhibition held in Brisbane, Australia, October 1618, 2000; C. A. Wright et al., “Downhole tiltmeter fracture mapping: A new tool for directly measuring hydraulic fracture dimensions,” SPE paper 49193, presented at 1998 SPE Annual Technical Conference, New Orleans, 1998; C. A. Wright et al., “Surface tiltmeter fracture mapping reaches new depths 10,000 feet, and beyond,” SPE paper 39919, presented at the 1998 SPE Rocky Mountain Regional Conference, Denver, Apr. 5-8, 1998; N. R. Warpinski et al., “Mapping hydraulic fracture growth and geometry using microseismic events detected by a wireline retrievable accelerometer array,” SPE paper 40014, presented at the 1998 SPE Gas Technology Symposium in Calgary, Canada, Mar. 15-16, 1998; R. L. Johnson Jr. and R. A. Woodroof Jr., “The Application of Hydraulic Fracturing Models in Conjunction with Tracer Surveys to Characterize and Optimize Fracture Treatments in the Brushy Canyon Formation, Southeastern New Mexico,” SPE paper 36470, presented at the 1996 Annual Technical Conference and Exhibition, Denver, Oct. 6-9, 1996; J. T. Rutledge and W. S. Phillips, “Hydraulic Stimulation of Natural Fractures as Revealed by Induced Microearthquakes, Carthage Cotton Valley Gas Field, East Texas,” Geophysics, 68:441-452, 2003; and N. R. Warpinski, S. L. Wolhart, and C. A. Wright, “Analysis and Prediction of Microseismicity Induced by Hydraulic Fracturing,” SPE Journal, pages 24-33, March 2004.
In at least one aspect, the disclosure relates to systems, apparatuses, and methods for neural network signal processing of microseismic events.
In at least one aspect, the disclosure relates to a method for neural network signal processing of microseismic events. The method can include disposing a series of sensors in at least a first well disposed adjacent to a second well. The method can also include obtaining a data signal measurement including noise events and microseismic acoustic emission events with the series of sensors. The method can include removing the noise events from the data signal measurement. The method can include determining with a first neural network an arrival time for each microseismic acoustic emission event.
In at least one aspect, the disclosure relates to a system for neural network signal processing of microseismic events. The system can include a series of sensors disposable in at least one first well positioned about a second well disposed in a subterranean formation. The series of sensors may obtain a data signal measurement including noise events and one or more microseismic acoustic emission events. The system can include a processor including a first neural network. The processor may remove the noise events from the data signal measurement and determine with the first neural network an arrival time for each microseismic acoustic emission event. The system can also include an interface that outputs the arrival time for each microseismic acoustic emission event.
In at least one aspect, the disclosure relates to a computer program product, including a computer usable medium with a computer readable program code embodied therein. The computer readable program code processes microseismic signal events, in that execution of the computer readable program code by one or more processors of a computer system causes the one or more processors to receive a data signal measurement of noise events and microseismic acoustic emission events from a series of sensors disposed in a first well. Execution of the computer readable program code may also cause the one or more processors to remove the noise events from the data signal measurement. Execution of the computer readable program code may also cause the one or more processors to determine with a first neural network an arrival time for each microseismic acoustic emission event.
This summary is provided to introduce a selection of concepts that are further described below in the detailed description. This summary is not intended to identify key or essential features of the claimed subject matter, nor is it intended to be used as an aid in limiting the scope of the claimed subject matter.
Embodiments of systems, apparatuses, and methods for neural network signal processing of microseismic events are described with reference to the following figures Like numbers are used throughout the figures to reference like features and components.
In the following description, numerous details are set forth to provide an understanding of the present disclosure. However, it will be understood by those skilled in the art that the present disclosure may be practiced without these details and that numerous variations or modifications from the described embodiments are possible.
Methods, systems and apparatuses presented herein are directed to signal processing to filter and automatically classify recorded microseismic events with a neural network based mapping technique. The signal processing may begin with filtering out background noise of a recorded signal. In an embodiment, the filtering of background noise may be performed with a wavelet based method, for example, as discussed further herein, or other suitable processing methods. The signal processing may also include identifying events present on various recorded waveforms (i.e., waveforms recorded on various channels). In an embodiment, the recorded signal has a relatively small amplitude due to the fact that the signal is based on microseismic events; thus, a high order statistic method can be employed to detect the events. A method of the present disclosure includes detecting small events in presence of colored noise, instead of white Gaussian noise. Because the number of events recorded can be vast, on the order of several thousands, a neural network based mapping can be applied to analyze and classify the recorded events in an automatic manner. In an embodiment, a plurality of monitoring wells, for example, two or more monitoring wells, may be used to record the data to train the neural network. In an embodiment, a monitoring well may be placed as a single spiral well about an injecting well.
A conventional method for monitoring acoustic emission (AE) events is to position an injection well and a monitoring well equipped with an array of sensors to listen. Since a time origin of each AE event is unknown, the formation speed information and the arrival time difference between different receivers can be used to invert for the location of each source event. Three component geophones can be used to narrow the direction of incoming AE waves. However, the acoustic emission is most accurately located if it occurs between two receivers, because the arrival time moves earlier in one receiver as the position of the event moving closer to it the corresponding arrival time moves later in time for the opposite receiver.
In a one dimensional example, the difference in the arriving time between two detectors with an AE event occurring in between the two detectors can be converted to the offset distance from the midpoint by multiplying the difference with the sound speed of the medium.
In a three dimensional example, the possible locations for the AE event will be limited to a hyperbola in between the monitoring well and the injection well. In using two wells (injection/monitor), however, even with three component geophones, microseismic events detection can be complicated. In order to improve reliability and efficiency of the conventional method, the present disclosure combines operational improvement (i.e., position and number of the wells) with an advanced signal processing technique.
Referring to
In one conventional method of monitoring, seismic events recorded in a single substantially vertical monitoring well can be subject to positional errors because of the time origin of each event is not known a priori, and the formation speed can vary as the acoustic event moves away from the borehole. An inverse problem of locating the origin of each microseismic event can be better constrained if the microseismic event occurs in between two distant detectors in two or more separate wells instead of a single well, and the velocity of the surrounding formation can be measured and established by cross-well survey beforehand. In an embodiment, it can be economical to drill a spiral well in which to dispose detectors rather than a plurality of wells.
In an embodiment, the method applies a Kohonen neural network, block 47 of
The different signals recorded by the receivers R1, R2, . . . RN (where N is the number of sensors) are filtered. Each signal recorded is a combination of the signal of interest with noise of various properties. The first stage includes filtering the recorded signal using an orthogonal wavelet transform, see, e.g., Ten Lectures on Wavelets, Ingrid Daubechies, SIAM: Society for Industrial and Applied Mathematics, 1992.
The algorithm to filter the recorded traces is now described, referring to the block 41 of
A signal x(t) is defined as a function of time t:
x(t)=f(t)+m(t)
where m(t) represents the noise corrupting the target data f(t), both as a function of time. The purpose of the processing is to filter the data in an automatic manner in order to extract the relevant information. An orthogonal wavelet basis reads:
{2−j/2φ(2−jt−k)}
with (j, k)ε2 and φ(t) is the mother wavelet and (j,k) represents the wavelet coefficient at a given scale. The wavelet coefficients of a discrete signal can be computed using a pyramidal algorithm described in Mallat, S., “Multi-resolution approximation and wavelet orthonormal bases of L2(R)”, Trans. Amer. Math. Soc., 315, 69-87, 1989. The wavelet coefficients of a discrete signal can enable substantially simultaneous examination of the information content of the analyzed signal in the time-scale half plane.
If the considered input signal, x(t), has a length of K coefficients, a filtered signal, Sf, will result by performing an inverse wavelet transform with a subset of the initial K wavelet coefficients, thereby automatically detecting the wavelet coefficients that may be used to reconstruct the denoised signal. In the case of white Gaussian noise, a filter criterion may be defined on a probability threshold ρ0 such as:
where Xn2 represents the Chi-square probability function with n degrees of freedom. The variance, σ2, of the noise may be a priori known. In an embodiment, the variance of the noise can be evaluated directly from the data. In an embodiment, the variance of the noise may be performed by assuming that the recorded information before the first arrival of interest will be representative of the noise. The probability threshold ρ0 establishes the level of risk that some noise may remain in the filtered signal. From Sf, the coefficients of the signal that will most accurately represent the denoised signal according to the criterion set above is such that
where wi corresponds to the wavelet coefficients. The sum is performed over the K-k wavelet coefficients discarded to reconstruct xF. The previous condition is achieved if xF(t) is constructed from the k wavelet coefficients with the largest coefficients cf Filtering non-stationary geophysical data with orthogonal wavelets, F. Moreau, D. Gibert. S, Saracco, Geophysical Research Letter, Vol 23, Issue 4, pages 407-410, 1996. When the coefficients are selected, an inverse wavelet transform can reproduce the denoised signal.
After the denoising is applied, a time delay estimation between two different sensors can be computed, as shown in block 42 of
To illustrate the high order statistic procedure, consider two signals, x(t) and y(t), having a delay of τ0. Noise signals, η(t) and m(t), are not correlated with the source wavelet but can be correlated together. The target data, f(t) and g(t), represent the impulse response of the medium, while s(t) represents the source signal:
x(t)=f(t)s(t)+m(t)
y(t)=g(t)s(t−τ0)+η(t)
In order to evaluate the time delay between the two signals, the correlation can involve assuming the nature of the noise. However, the noise can be greater than the target data signal. As such, a bicoherence-correlation approach can be used to evaluate the delay between the two sensors. A bicoherence, BC, can be defined as:
where ω1, ω2 correspond to frequencies
with
where E corresponds to an expectation function, and P, the power spectrum, is defined as:
P
xy(ω)=E[X*(ω)Y(ω)]
where X(ω) and Y(ω) represent a Fourier transform of x(t) and y(t) and * represents a complex conjugate.
A bicoherence correlation (BCC) can be defined as:
where TF−1 corresponds to an inverse Fourier transform. The delay between the two signals will be indicated by the maximum peak of the bicoherence correlation function, see, e.g., Yung, S. K., and Ikelle, L. T., 1997, An example of seismic time picking by third order bicoherence: Geophysics, 62, 1947-1951.
The result of the time delay estimation may be used, in turn, to remove signals that have relatively large absolute differential delay, as shown in block 43 of
p
m=median{pk}
where a median absolute deviation (MAD) scale estimator can be computed as:
P
MAD=1.4826median{|pk−pm|}
The value of median {|pk−pm|} is a measure of how far the data point pk typically lies from the reference value pm. A normalization factor 1.4826 is based on the fact that a nominal part of the data sequence {pk} has a Gaussian distribution. PMAD represents an unbiased estimator of a standard deviation σ when normalized in this manner. A studentized deviations may be calculated as follows:
In an embodiment, a point may be identified as an “outlier” if |zk|>ξ where ξ represents a threshold value. In an example embodiment, ξ may be set to 3, but can be changed without departing from the scope of this disclosure. The procedure can be implemented at a well site with minimal computational resources. Other suitable techniques to detect outliers may be substituted for the procedure described above without departing from the scope of this disclosure.
Turning now to block 44 of
In principle, the robust PCA estimation involves replacing a standard estimation of the covariance matrix with a robust estimator of the covariance matrix. In order to do so, an expectation maximization algorithm can be used, as described for example, in F. Ruymagaart, “A Robust Principal Component Analysis”, J. Multivariate Anal., Vol. 11, pp. 485-497, 1981; and M. Tipping and C. Bishop, “Probabilistic principal component analysis”, Journal of the Royal Statistical Society B, 61, 611-622, 1999.
In an embodiment, the number of components used to represent the signal subspace may be variable and can vary from one data set to another. In an example for purposes of illustration, the number of components kept to represent the signal subset can be defined as:
where λi represents the ith eigenvalue of the data matrix, and where p and q are defined by either 1) a plot of the eigenvalues, λi, as a function of i, or 2) a predetermined threshold of the percentage of energy represented in the reconstructed data, and r represents the total number of eigenvalues coming from the covariance matrix computed from the data. This criterion will give the percentage of the energy, which is contained in the reconstructed signal. In block 45 of
Turning to block 46 of
In block 48 of
With adequate external information, modeling of previous experiments in similar environments, it is possible to have enough data to train the neural network. The effectiveness of the neural network may be based on the amount of data set used for training. The more used, the better the results will be. In an embodiment, three or more data sets should be sufficient. In an embodiment, a database could contain the data of many measurements and/or experiments previously performed for use in a training phase for neural network processes. In an embodiment, the training of the neural network as well as event classification by the neural network will be performed in the time-frequency domain for the reasons mentioned above.
Without adequate external information, the training phase for the neural network can be more complicated. In this case, one or more sensors can be randomly selected, the measurements from which will be used for training the neural network. In an embodiment, the “training sensors” can be used repeatedly during the training phase for the neural network.
To summarize, noises detected in multiple channels are separated from the events of interest and subsequently removed using a combination of covariance analysis (block 43), principal component analysis (block 44), and differential time delay estimates (block 42). A self-organizing map is then applied in a trained neural network (trained in block 47) to separate the AE's from noise in the remaining data (block 48). At the point in the method in which the number of true AE events have been sorted out from noise, the number of events may be manageable to be processed individually by a person. In an embodiment, an additional neural network, independent from the first, can be used to locate each AE event. For example, the second additional neural network may be of the form used to identify source locations, where numerical simulation data based on known source locations are used to train and optimize the weights used in the neural network.
As those skilled in the art will understand, one or more of the steps of methods discussed above may be combined and/or the order of some operations may be changed. Further, some operations in methods may be combined with aspects of other example embodiments disclosed herein, and/or the order of some operations may be changed. The process of measurement, its interpretation and actions taken by operators may be done in an iterative fashion; this concept is applicable to the methods discussed herein. Finally, portions of methods may be performed by any suitable techniques, including on an automated or semi-automated basis on computing system 500 in
Portions of methods described above may be implemented in a computer system 500, one of which is shown in
In one implementation, petroleum real-time data from the sensors may be stored in disk storage device 531. Various non-real-time data from different sources may be stored in disk storage device 533. The system computer 530 may retrieve the appropriate data from the disk storage devices 531 or 533 to process data according to program instructions that correspond to implementations of various techniques described herein. The program instructions may be written in a computer programming language, such as C++, Java and the like. The program instructions may be stored in a computer-readable medium, such as program disk storage device 535. Such computer-readable media may include computer storage media. Computer storage media may include volatile and non-volatile, and removable and non-removable media implemented in any suitable method or technology for storage of information, such as computer-readable instructions, data structures, program modules or other data. Computer storage media may further include RAM (Random Access Memory), ROM (Read Only Memory), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other solid state memory technology, CD-ROM, digital versatile disks (DVD), or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other suitable medium which can be used to store the desired information and which can be accessed by the system computer 530. Combinations of any of the above may also be included within the scope of computer readable media.
In one implementation, the system computer 530 may present output primarily onto graphics display 527, or via a printer (not shown). The output from computer 530 may also be used to control instruments within the steam injection operation. The system computer 530 may store the results of the methods described above on disk storage 529, for later use and further analysis. The keyboard 526 and the pointing device (e.g., a mouse, trackball, or the like) 525 may be provided with the system computer 530 to enable interactive operation.
The system computer 530 may be located on-site near the well or at a data center remote from the field. The system computer 530 may be in communication with equipment on site to receive data of various measurements. Such data, after conventional formatting and other initial processing, may be stored by the system computer 530 as digital data in the disk storage 531 or 533 for subsequent retrieval and processing in the manner described above. While
Although a few example embodiments have been described in detail above, those skilled in the art will readily appreciate that many modifications are possible in the example embodiments without materially departing from this disclosure. Accordingly, such modifications are intended to be included within the scope of this disclosure as defined in the following claims. In the claims, means-plus-function clauses are intended to cover the structures described herein as performing the recited function and not simply structural equivalents, but also equivalent structures. Thus, although a nail and a screw may not be structural equivalents in that a nail employs a cylindrical surface to secure wooden parts together, whereas a screw employs a helical surface, in the environment of fastening wooden parts, a nail and a screw may be equivalent structures. It is the express intention of the applicant not to invoke 35 U.S.C. §112, paragraph 6 for any limitations of any of the claims herein, except for those in which the claim expressly uses the words ‘means for’ together with an associated function.
